... slightly by having indepen-dent parameters for 1-count, 2-count, and many-count n-grams, but still assumes that¯d(i) is constant for i greater than two. Second, by using the samediscount for ... for a given n-gram countis well-approximated by its mean. For similar cor-pora, this seems to be true, with a histogram of testcounts for trigrams of count 10 that is nearly sym-metric. For ... interpolation parameters for each order. Param-eters for GDLM, MKNLM, and KNLM are initial-ized based on estimates from¯d(i): the regressionthereof for GDLM, and raw discounts for MKNLMand KNLM....